automatic generation of design structure matrices through the … · 2017-11-28 · faculty of...

General rights Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights.

Users may download and print one copy of any publication from the public portal for the purpose of private study or research.

You may not further distribute the material or use it for any profit-making activity or commercial gain

You may freely distribute the URL identifying the publication in the public portal If you believe that this document breaches copyright please contact us providing details, and we will remove access to the work immediately and investigate your claim.

Downloaded from orbit.dtu.dk on: Apr 08, 2020

Automatic generation of design structure matrices through the evolution of productmodels

Gopsill, James A.; Snider, Chris; McMahon, Chris; Hicks, Ben

Published in:Artificial Intelligence for Engineering Design, Analysis and Manufacturing

Link to article, DOI:10.1017/S0890060416000391

Publication date:2016

Document VersionPublisher's PDF, also known as Version of record

Link back to DTU Orbit

Citation (APA):Gopsill, J. A., Snider, C., McMahon, C., & Hicks, B. (2016). Automatic generation of design structure matricesthrough the evolution of product models. Artificial Intelligence for Engineering Design, Analysis andManufacturing, 30(4), 424–445. https://doi.org/10.1017/S0890060416000391

https://doi.org/10.1017/S0890060416000391

https://orbit.dtu.dk/en/publications/6c214645-bbd2-45e9-bc39-20236bdaead0

https://doi.org/10.1017/S0890060416000391

Automatic generation of design structure matrices through theevolution of product models

JAMES A. GOPSILL, CHRIS SNIDER, CHRIS MCMAHON, AND BEN HICKSFaculty of Engineering, University of Bristol, Bristol, United Kingdom

(RECEIVED October 1, 2015; ACCEPTED May 31, 2016)

Abstract

Dealing with component interactions and dependencies remains a core and fundamental aspect of engineering, where con-flicts and constraints are solved on an almost daily basis. Failure to consider these interactions and dependencies can lead tocostly overruns, failure to meet requirements, and lengthy redesigns. Thus, the management and monitoring of these de-pendencies remains a crucial activity in engineering projects and is becoming ever more challenging with the increasein the number of components, component interactions, and component dependencies, in both a structural and a functionalsense. For these reasons, tools and methods to support the identification and monitoring of component interactions and de-pendencies continues to be an active area of research. In particular, design structure matrices (DSMs) have been extensivelyapplied to identify and visualize product and organizational architectures across a number of engineering disciplines. How-ever, the process of generating these DSMs has primarily used surveys, structured interviews, and/or meetings with en-gineers. As a consequence, there is a high cost associated with engineers’ time alongside the requirement to continuallyupdate the DSM structure as a product develops. It follows that the proposition of this paper is to investigate whether anautomated and continuously evolving DSM can be generated by monitoring the changes in the digital models that representthe product. This includes models that are generated from computer-aided design, finite element analysis, and computa-tional fluid dynamics systems. The paper shows that a DSM generated from the changes in the product models corroborateswith the product architecture as defined by the engineers and results from previous DSM studies. In addition, further levelsof product architecture dependency were also identified. A particular affordance of automatically generating DSMs is theability to continually generate DSMs throughout the project. This paper demonstrates the opportunity for project managersto monitor emerging product dependencies alongside changes in modes of working between the engineers. The applicationof this technique could be used to support existing product life cycle change management solutions, cross-company productdevelopment, and small to medium enterprises who do not have a product life cycle management solution.

Keywords: Computer-Aided Design; Design Structure Matrices; Graph Theory; Metadata; Network Analysis

1. INTRODUCTION

Handling component interactions and dependencies remainsa core and fundamental activity, with engineers resolvingconflicts and satisfying constraints on an almost daily basis.Failure to consider these interactions and dependencies canlead to costly overruns, failure to meet requirements, productrecalls, and/or lengthy redesigns. For example, issues in thelength of cabling required in the A380 led to a $6.1 billion de-lay for the project, while Toyota has had to recall 625,000 ve-hicles because of faulty hybrid software (Calleam, 2011;Bruce, 2015).

In each of these examples, the companies were introducinginnovative technologies to their product lines, which led tothese complex products containing many more componentsthat are more highly interconnected and dependent on one an-other, structurally, behaviorally, and functionally (Hamraz &Clarkson, 2015). This is further supported by Briggs (2012),who details that the Boeing 787 Dreamliner consists of over300,000 parts modeled in computer-aided design (CAD),and the accompanying product data management systemcommonly saw between 75,000 and 100,000 accesses aweek. This highlights the scale of CAD work being per-formed by a company that has locations across the globeand employs approximately 160,000 people. Due to this scaleand the accompanying complexity, the ability to monitor andevaluate the interactions and dependencies between compo-

Reprint requests to: James A. Gopsill, Faculty of Engineering, Universityof Bristol, 0.36 Queen’s Building, University Walk, Bristol BS8 1TR, UK.E-mail: [email protected]

Artificial Intelligence for Engineering Design, Analysis and Manufacturing (2016), 30, 424–445.# Cambridge University Press 2016 0890-0604/16 This is an Open Access article, distributed under the terms of the Creative Commons Attribution licence(http://creativecommons.org/licenses/by/4.0/), which permits unrestricted re-use, distribution, and reproduction in any medium, provided the original work isproperly cited.doi:10.1017/S0890060416000391

424

https://www.cambridge.org/core/terms. https://doi.org/10.1017/S0890060416000391Downloaded from https://www.cambridge.org/core. DTU Library - Tech Info Ctr of Denmark, on 22 Nov 2017 at 13:33:32, subject to the Cambridge Core terms of use, available at

mailto:[email protected]

https://www.cambridge.org/core/terms

https://doi.org/10.1017/S0890060416000391

https://www.cambridge.org/core

nents at various system and subsystem levels remains a crit-ical activity.

For these reasons, the development of supportive toolsand methods continues to be an active area of research,with design structure matrices (DSMs) emerging as a com-mon method. Originally conceived in the 1980s by Steward(1981) as a branch of graph theory, DSM seeks to understandthe connected nature of engineering systems through a N�Nmatrix of interactions between system elements (Eppinger,1997). The system elements can consist of individual compo-nents, assemblies of components, systems of components, en-gineers, teams of engineers, processes, and/or organizationalstructure to name a few. DSMs enable researchers to analyzeand visualize product, process, organisational, and multido-main architectures of engineering products and projects.

Over the last 20 years, DSM analysis has been shown to pro-vide insights into the product architecture of engineeringproducts as well as the design process, and has helped increaseengineering companies’ understanding of their product (see, fore.g., Pimmler & Eppinger, 1994; MacCormack et al., 2006;Sosa et al., 2007; Jarratt et al., 2011). This understanding pro-vides the basis for companies to better manage change prop-agation and risk, model change scenarios, and increase themodularity of their product designs. In addition, analysis oforganizational architectures has resulted in companies beingable to increase the performance of their development teamsthrough restructuring in accordance with the product architec-ture (Eppinger, 1997; Browning, 2009). DSM has also beensuccessfully applied to understand and evaluate supply net-works as well as in integrating multiple complex systemssuch as the internal combustion engine and electric recovery

and deployment system for hybrid vehicles (Fixson, 2005;Chen & Huang, 2007; Gorbea et al., 2008).

In order to generate DSMs, it is first necessary to under-stand the interactions between the system elements of interest.To capture these interactions, the majority of reported studieshave used surveys, manual coding of engineering documents,structured interviews, and/or meetings with engineers (Fig. 1).In addition, many of the reported studies often use binary orordinal values to determine the level and/or type of depen-dency between system elements. For example, Sosa et al.(2003) used a range from –2 toþ2 in order to provide a weigh-ted degree of dependency between system elements of a jet en-gine, while Gorbea et al. (2008) used a binary value to indicatethe presence of dependencies between the component andfunction domains of hybrid vehicles. Although the captureof these data types has been shown to provide insightful resultsin terms of the dependencies within product architectures, it isargued that they may also be a limiting factor in providingeven greater insights due to the granularity and resolution inwhich system dependencies can be assessed.

Because these data capture methods involve considerable in-teraction with engineers, there is inherently a cost associatedwith the engineer’s time, potential error due to the subjectivityand precision of recall of the responses from engineers, andlimitations in terms of determining the strength of the depen-dencies. Due to the time commitment required, data gatheringis often a discrete activity that occurs at a single point in theproject. Hence, a DSM can be considered a snapshot of theproduct architecture. For these reasons, it is contended thatthere is a need to explore alternative means to generateDSMs and, in particular, means that are automated, more ob-

Fig. 1. Data collection methods from the studies in Eppinger and Browning (2012).

Automatic generation of design structure matrices 425



https://doi.org/10.1017/S0890060416000391


jective and real time, and that could support or supplement ex-isting manual elicitation strategies (Senescu et al., 2012).

In the case of mature products such as aircraft and motorvehicles, the product architecture is generally well understoodand the majority of dependencies are known. These are oftencaptured through a combination of process models, productdata management, product life cycle management (PLM),and associative CAD systems (Jarratt et al., 2011). However,such a proactive approach is not often possible for new, lowvalue, or one-time large complex products/systems. It isalso argued that an automated and real-time DSM analysiscould provide additional support to existing solutions.

Many cases remain where a PLM change management so-lution may not be viable such as in small to medium enter-prises where the expense may be too great. In addition, cross-company collaborative engineering projects can often pro-vide additional challenges in monitoring dependencies dueto the introduction of multiple disparate enterprise informa-tion technology systems having to interface within one an-other as well as issues with data security.

Therefore, the proposition for this paper is to investigatewhether an automated and continuously evolving DSM canbe generated from the changes to the digital models thatrepresent the product. Such models include CAD, finite elementanalysis (FEA), and computational fluid dynamics (CFD). Toexplore this proposition, this paper presents the results from aDSM analysis of the evolution of product models for a formulastudent team. The objectives of this paper are threefold:

1. automatically identify component and subsystem de-pendencies across CAD models,

2. automatically identify cross-disciplinary dependenciesthrough the analysis of the full range of product models,and

3. monitor the evolution of dependencies within theproduct models.

The paper continues by providing details of the method ofdata capture and the process by which DSMs have been auto-matically generated. The process uses the co-occurrence ofproduct model edits to elicit potential candidates for depen-dency. This is then evaluated through the creation of aCAD, product model, and dynamic product model DSMs.The results of these have been verified through comparisonwith the product architecture as defined by the design teamand results from previous DSM studies. In addition, particularattention is given to the affordances offered by dynamicproduct model DSMs. The paper then concludes by high-lighting the key findings from the DSM analysis and areasof future work.

2. CAPTURING THE EVOLUTION OF PRODUCTMODELS

The data set used in this study has been generated from theevolution of product models produced by the Formula Stu-

dent team at the University of Bath. Formula Student (alsoknown as Formula SAE) is a motor-sport educational pro-gram whereby teams of students from competing universitiescreate a single-seat race car that competes in various chal-lenges set out by the competition organizers. The competi-tions are held worldwide and include events in the UnitedKingdom, United States, Australia, and Europe.

The team consisted of 33 engineering students in their finalyear of study who have undertaken a range of engineeringcourses, including automotive, aerospace, electrical, manu-facturing, and mechanical. Therefore, it is argued that this re-flects a highly multidisciplinary colocated engineering proj-ect environment.

In addition, the team have developed a bespoke lightweightCAD management tool that manages the naming convention,relationships, and organization of the CAD models. This isknown as the bath automated parts system (BAPs). Allwork on the models is performed on the shared network drive.

In order to monitor the evolution of the product models, aRaspberry Pi (connected to the network) was used to monitorthe status of the shared network drive at 20-min intervals(Raspberry Pi Foundation, 2015). More specifically, the folderstructure alongside the metadata attributes of all the productmodels was captured. This included model size, date accessed,and date modified. The data capture was performed over a13-week period from March 2014 to June 2014.

Table 1 summarizes the properties of the data set with re-spect to the various product models for the Formula Studentcar. CAD models total 1432 models, which have been up-

Table 1. Summary of the formula studentengineering models

Description Value

Weeks of data capture 13Total CAD models created 1,432Total CAD models assigned within BAPs 541

Brake system 13Complete assembly 1Electrical 9Engine & drivetrain 142Frame & body 145Miscellaneous 1Standard parts 45Steering 46Suspension 113Wheels, wheel bearings, & tires 26Unassigned CAD models 891

CAD model updates 10,145CFD models 227CFD updates 681FEA models 43FEA updates 121WAVE models 95WAVE model updates 245OptimumK models 44OptimumK model updates 88

J.A. Gopsill et al.426



https://doi.org/10.1017/S0890060416000391


dated 10,145 times using the Siemens NX CAD package(Siemens, 2015). Table 2 shows that the BAPs CAD manage-ment tool classification covers 42% of the CAD models gen-erated. Fifty-eight percent remains formally unclassified, andsuggests that there is a high volume of temporary models aswell as models that support other aspects of the design pro-cess, such as FEA and CFD modeling.

The second most active area in terms of product model up-dates are models generated from ANSYS CFX, with 227unique simulation setup models and 681 updates (ANSYS,2015). This is followed by the Ricardo WAVE one-dimen-sional gas dynamics modeling, which models the behaviorof the internal combustion engine and has 95 simulation setupmodels and 245 updates (Ricardo, 2015). Last in terms of rel-ative model activity (number of updates) is the FEA and Op-timumK (a dynamics simulation package) modeling, whichsaw 43 and 44 unique models with 121 and 245 total updates,respectively.

3. DSM GENERATION

In general, the generation of DSMs has involved the elicitationof dependencies between systems and subsystems through ex-pert opinion. In this paper, it is posited that models editedwithin a certain time period of each other can be consideredto be candidates for a potential dependency. The probabilityof this dependency being true, alongside the potential strengthof this dependency, is increased by a consistent reoccurrenceof the models being edited within defined time periods.

An initial proof-of-concept investigation was conducted bySenescu et al. (2012), who sought to generate DSMs throughthe writing, viewing, and exporting of engineering docu-ments within a cloud-based information management tool.The investigation concluded that dependencies were able tobe identified; however, a significant amount of false positiveswere also identified due to pure chance of a document changeco-occurring and/or the concurrency of work within the proj-ect. Therefore, this paper builds upon these results and pro-poses a seven-stage process for generating DSMs throughthe analysis of the co-occurrence of model edits, with eachstage featuring techniques to reduce the false positive identi-fication within the DSM (Fig. 2).

Stage 1 involves selecting the appropriate models that willbe used as the basis for the DSM and is based on the level of

activity recorded. The process then enters an iteration loopencompassing Stages 2 to 6, where a number of potentialDSMs are generated by varying the time period in whichco-occurrences can occur and the level of pruning. Pruningrefers to the removal of edges (dependencies in this case)that are of a low weighting and is a method of removing noisefrom the matrix. Stage 2 involves generating the dependencymatrix through the co-occurrence of model activity. In thefirst instance, it has been assumed that the matrix is directed(i.e., if Model A changes, then Model B changes). Stage 3then examines this assumption and assesses the “directed-ness” of the matrix through pairwise comparison of the di-rected matrix elements (i.e., Model A to B and Model B toA). A decision is then made as to whether to continue theanalysis with a directed or undirected matrix. This leadsinto Stage 4, where the matrix is weighted. The objective ofthe weighting is to evaluate the likelihood of a dependencywhile attempting to filter potential false positive dependen-cies introduced by the method of measuring co-occurrencethrough model edits. Once the weighting scheme has been ap-plied to the matrix, pruning of the matrix can occur (Stage 5).This is particularly important as the level of pruning cangreatly affect the partitioning of the matrix and a compromisemust be sought because excessive pruning of the matrix mayresult in models becoming isolated. After the pruning, it is thena case of partitioning the matrix to reveal the structure andmodels with a high likelihood of being dependent (Stage 6).Stage 7 of the process then evaluates the generated DSMs interms of the modularity, number of partitions, number of com-ponents, and remaining matrix size in order to determine themost appropriate time period and level of pruning for the givendata set. This also reveals the sensitivity of the analysis with re-spect to the level of pruning and time period at which a co-oc-currence is detected. This section continues by discussing eachstage of the process in further detail alongside an example ofthe generation of the CAD DSM presented in this paper.

3.1. Initial model selection

The first stage of the process requires the specification of themodels to be analyzed. In this study, the CAD, CFD, FEA,WAVE, and OptimumK models were specified as potentialcandidates for the DSM. This is achieved by detecting the re-spective file extensions for the product models. In addition,the number of updates for each model is also used as a filter,and Figure 3 shows the distribution of model activity for themodels in terms of the number of changes made to the modeland the number of days the model has been actively workedupon (i.e., an update has to have been made on the day forit to be counted). It can be seen that there is a slight positivecorrelation with the number of days the model has been ac-tively worked upon and the number of edits to the model.In addition, one model in particular appears to exist outsideof the main group of model activity. Upon inspection of thefile name, it was revealed that this model embodies the fullCAD assembly of the car, and it is likely that the number of

Table 2. Reduction in models through filteringbased on level of edit activity

Model Type Filtered Nonfiltered

CAD (assigned) 296 541CAD (unassigned) 17 891CFD 7 227FEA 3 43OptimumK 0 7




https://doi.org/10.1017/S0890060416000391


Fig. 2. Process of automatically generating design structure matrices.

Fig. 3. Distribution of product model activity.




https://doi.org/10.1017/S0890060416000391


updates to this model reflects the number of model updatesarising from the various subassemblies of the product.

A closer inspection of the main groups model activity(Fig. 3) reveals that the CAD models embody a much greateractivity and days of activity compared to the CFD, FEA, andWAVE models. This provides evidence to suggest that theCAD model plays a fundamental role in the Formula Studentproject with the CFD, FEA, and WAVE models used to sup-port the development of the product rather than lead the de-sign of the product.

To continue the analysis further, this stage requires the userto decide upon a suitable level of co-occurred activity given thecontext. In this case, models with an activity of greater thanfour updates have been selected. This includes the creation fol-lowed by three more updates. The aim is to remove the modelsthat show little to no activity and, thus, may represent areas ofthe product that cannot be altered by the team (e.g., a bought inpart such as the engine block and/or gearbox). Imposing a lowthreshold for the number of updates helps prevent an overre-duction in the size of data set. For the case of the CAD models,the data set consists of 541 formally classified models reduced(cf. BAPs; Table 2) to 296 (55%) for further analysis while theunclassified CAD models are reduced from 891 to 17.

3.2. Co-occurrence matrix generation

Once the filtered data set has been defined, the generation ofthe DSM matrix can commence. This is achieved through theanalysis of the co-occurrence of product model changeswithin the filtered data set. Figure 4 shows the process ofidentifying the co-occurrence of model changes. TakingModel A as the model of interest, the process identifies allthe timestamps where the model has changed and generatesa period in time where the occurrence of changes to and crea-tion of other models are noted. Thus, if Model B has changedwithin this period, then this is defined as a co-occurrence andis a potential candidate for a dependency. The more often thatthis occurs, the greater the likelihood of the existence of a de-pendency. Currently, the impact of a change and the creation

of a new model within the time period are treated equally. Themaximum time period is defined by the useralthough the periodcan be shorter if model A changes again within its own timeperiod as demonstrated by the 4, 3, and 2 h in Figure 4.

Figure 4 also shows the three possible scenarios that canoccur. Scenario (i) occurs where Model A has changed andwithin the time period, Model B has also changed. A co-oc-currence of model activity is consequentially recorded. Sce-nario (ii) occurs where Model B changes again within thesame time period. In this scenario, a co-occurrence has al-ready been recorded, and therefore, further changes are not re-corded because they exist within the same time period. Sce-nario (iii) occurs where Model B changes within a periodof inactivity of Model A. In this case, there is no co-occur-rence of model activity.

The analysis of co-occurrence results in a DSM as illus-trated in Figure 5, where a change in the model in the rowvector has a likelihood of leading to a change in the modelin the column vector. As a result of the generation process,the DSM is inherently directed where Model A can result ina change in Model B, but Model B may not lead to a changein Model A. This forms the initial matrix that, in the exam-ple of the CAD DSM with a max time period of 1 h, has ledto a DSM consisting of 296 models and 32,508 directed co-occurrences ranging from 1 to 77 in weight.

In creating this matrix, there is an inherent assumption thatthe dependencies are directed, and thus, there is a need to as-sess the truth of this assumption and whether an undirectedmatrix could be equally effective in representing the dataset (Stage 3). There is currently a wider range of validatedpartitioning algorithms in relation to undirected matrices,while directed partitioning algorithms remains a maturingfield of research. Therefore, it is argued that, where possible,an undirected matrix would be a preferred result.

In addition, with many co-occurrences having a weight of 1also highlights the likelihood that a number of false positivedependencies have been captured. Therefore, there is a needto weight the resulting undirected/directed matrix so as to re-duce the noise and help reveal key co-occurrences (Stage 4).

Fig. 4. Determining co-occurrence of model activity.




https://doi.org/10.1017/S0890060416000391


3.3. Evaluating the “directedness” of the matrix

Stage 3 assesses the directedness of the generated matrix.This involves examining the number of one-way co-occur-rences (10,850; i.e., a co-occurrence has been indicatedfrom Model A to B but not from Model B to A) and the num-ber of two-way co-occurrences (10,829; i.e., co-occurrenceshave been detected for both directions from A to B andfrom Model B to A). This is examined through the pairwisecomparison of the two-way co-occurrences (i.e., the weight-ing of the co-occurrence from Model A to B and B to A)and the distribution of the one-way co-occurrence values.The results are shown in Figure 6. The pairwise comparison

of the two-way co-occurrences (Fig. 6a) reveals that there islittle difference between the co-occurrence values. It can beseen that 72% of all the two-way co-occurrences have a dif-ference of less than 2 (e.g., if the co-occurrence value forModel A to B is 10, then the co-occurrence value for ModelB to A would be 8 or 12 for a difference to be 2). Similarly,Figure 6b shows that the one-way co-occurrence values haverelatively low co-occurrence values, with 77% of one-way co-occurrences also having a value less than 2 (e.g., if the co-oc-currence value for Model A to B is 2 with no co-occurrencedetected in the direction of Model B to A). In both cases, it isargued that there is little directionality within the matrix, and

Fig. 5. Illustration of the design structure matrix from the co-occurrence of model activity.

Fig. 6. Analyzing directedness of the computer-aided design structure matrices matrix.




https://doi.org/10.1017/S0890060416000391


the one-way co-occurrences are likely to be false positive de-pendencies because they have low co-occurrence values.

From these results, it is argued that an undirected matrix canbe equally representative of the data as a directed matrix.Therefore, all DSMs in this paper have been translated from di-rected to undirected networks prior to weighting by combiningthe number of co-occurrences for both directions (i.e., the sumof the weights from Model A to B and Model B to A).

3.4. Matrix weighting

In order to further reduce the likelihood of false positives andreveal the strongest candidates for dependencies, a matrixweighting scheme has been applied. The aim of the schemeis to reduce the potential influence of two or more designersmodifying one or more unrelated documents at the same time.Two schemes are available, and the selection depends onwhether a directed or undirected matrix has been generatedfrom Stage 4.

For a directed matrix, each row vector is normalized by di-viding the vector by the total number of changes made to themodel that the row represents; that is, the co-occurrence be-tween Model A and Model B would be divided by the totalnumber of changes to Model A. This gives the ratio of co-oc-currence between the two models in relation to the total modelactivity of Model A.

For an undirected matrix, each cell is divided by the sum ofthe total number of changes made to both models that the cellrepresents, that is, the number of co-occurrences betweenModel A and B divided by the sum of the total number tochanges made to Models A and B. The premise is that higherratios highlight a greater likelihood of the existence of a depen-dency between the two models. The weighting scheme alsonormalizes the co-occurrence of models and, thus, enablescomparison between the range of model activities observed.

In the case of the CAD DSM, the undirected weightingscheme has been applied, and Figure 7 shows the impact ofthe weighting scheme on the distribution of the model co-oc-currence values. It is immediately apparent that the weightingscheme increases the relative importance of some of the co-occurrences, while in the original matrix, there is a sharpdrop off due to the large variability in the model activityacross the CAD models. Consequentially, it is argued thatthe weighting scheme has potentially highlighted more ofthe positive dependencies.

3.5. Matrix pruning

Pruning is where the co-occurrences of given weights are re-moved from the DSM. In graph theory, this would be the re-moval of edges of a certain weighting from a network. Thekey objective of the pruning is to reduce the false positivesin the matrix while maintaining the structure of the matrixand, in this case, revealing the candidate dependencies ofhigher likelihood. In this case, co-occurrences of a lowweighting are removed because it is deemed that these arefalse positive candidates for dependency. The appropriatevalue at which to prune is determined later in Stage 7, wherean optimization takes place in order to determine the most ap-propriate values for the time period and level of pruning.

Pruning the matrix also has the ability to disconnect mod-els and groups of models from the rest of the matrix. In thefirst case, the process removes this model from the rest ofthe analysis, while in the second case, the group of models re-main within the matrix as a separate group. These are com-monly referred to as components in graph theory and canbe considered to be a form of partitioning for the DSM.Even with a matrix that has been separated into a numberof components, further partitioning can take place in orderto uncover partitions within the components that remain.

Fig. 7. Distribution of weighted co-occurrences within the computer-aided design structure matrices.




https://doi.org/10.1017/S0890060416000391


In the case of the CAD DSM presented in this paper, thepruning level was set at 0.25, and this resulted in the matrixbeing reduced from 296 to 253 models due to disconnectionand the number of co-occurrences being reduced from 21,679to 613. The significant reduction in the number of co-occur-rences further highlights the level of false positive that are in-troduced using this technique. In addition, the high number ofremaining models in the matrix after pruning (85.4%) showsthat the 613 remaining co-occurrences have the potential toreflect the main structural elements of the matrix. In addition,the pruning led to the creation of 13 components within thematrix. Components are akin to modules within engineeringproducts, where the models within the component are highlyconnected and dependent from the rest of the product archi-tecture. The number of components can be considered to bean indicator of how modular a product design is.

3.6. Matrix partitioning

Because the formation of the matrix involves the measure-ment of the co-occurrence of model activity and correspond-ingly results in potential false positive dependencies appear-ing, it is argued that typical DSM partitioning techniques maynot be as suitable as other graph theory partitioning tech-niques. This is because DSM partitioning techniques areusually performed on data that is typically binary or ordinal,and there is an assumed confidence in the values attainedfrom data capture because it is often through expert opinion.

In contrast, the co-occurrence measurements in this analy-sis are transactional and continuous, as well as having achance of some of the candidate dependencies being false pos-itives. Therefore, the structure of the data can be considered tobe more akin to communication transactions such as e-mailand social media (Blondel et al., 2008; Pujol et al., 2009). Inthis field, the Louvain community-partitioning algorithm hasbeen shown to perform well in identifying partitions of com-munities. For these reasons, the Louvain method has beenadopted in this DSM generation process.

The Louvain community algorithms’ objective is to generatea set of partitions for the matrix that returns the highest modu-larity value. Modularity (Q) is an assessment of the quality ofthe matrix partition and is defined as (Newman, 2004):

Q ¼ 12m

Xij

Aij �kikj

2m

� �@(ci, cj), (1)

where m ¼ ð1=2ÞP

ij Aij

� �and is the number of co-occurrences

within the matrix; @ is the Kronecker delta function, which is 1if a co-occurrence exists between two models and 0 otherwise;(kikj)/2m is the probability that a co-occurrence may exist be-tween two models, where ki is the number of models thathave co-occurrences with model i and ki is the number ofmodels that have co-occurrences with model j; and Aij is theweighted co-occurrence between the two models in the matrix.

In order to obtain the highest modularity partition, the al-gorithm iterates between two steps. The first assigns each

model to its own partition. This is then followed by the algo-rithm sequentially moving one model to a different partitionand calculating the change in modularity. From this, the max-imum modularity change can be identified.

The second step merges the models together to form a par-tition of models and combines the co-occurrences of modelactivity to form single co-occurrence links to the rest of thematrix, and self-loops are used to identify internal co-occur-rences between the models within the partition. The aim is toachieve a partitioning whereby each partition is highly con-nected internally and weakly connected to one another.

Thus, it can be considered a form of hierarchical clustering,and the algorithm iterates until the modularity can no longerbe increased by partitioning the models. This paper uses thecommunity application programming interface implementa-tion of the Louvain community-partitioning algorithm withinthe NetworkX python package (Hagberg et al., 2008).

In the case of the CAD DSM, with a time period of 1 h andpruning of 0.25, the partitioning achieves a modularity scoreof 0.851 with 26 partitions being generated. This value iscomparatively high when compared to many other networkdata sets where a modularity greater than 0.3 has been consid-ered to be a “good” level of partitioning (Newman, 2004). Al-though modularity has been typically been used as an indica-tor of a “good” partition, it cannot provide completeconfidence because random networks are able to attain arange of modularity scores (Newman, 2006). In this study,the high score of 0.851 is expected given the typical hierarchi-cal nature of product architectures that often have well-de-fined system and subsystem boundaries.

3.7. Sensitivity of DSM generation to time period andpruning values

As previously stated, one of the main challenges in automat-ically generating DSMs through the co-occurrence of modelchanges is determining the appropriate values for the time pe-riod in which co-occurrences are detected and the level ofpruning. This is unlike a normal DSM approach, where thereis an assumed confidence in the validity of the dependenciesbecause they are typically elicited from a number of experts.The process therefore incorporates a stage to determine themost appropriate values and assess the sensitivity of theDSM to these parameters. This involves optimizing the ma-trix with respect to the following:

1. the quality of the partitioning achieved (i.e., a “good”system and subsystem product architecture can be iden-tified),

2. the number of partitions detected (i.e., the level of sub-system granularity that can be attained),

3. the remaining size of the matrix after pruning (i.e., howrepresentative the DSM is to the entire product architec-ture), and

4. the number of matrix components (i.e., how modularthe product design is).




https://doi.org/10.1017/S0890060416000391


First, the quality of the partitioning achieved by the processis assessed by taking the modularity into consideration wherea higher value is preferable because this highlights that theanalysis is able to identify clear system and subsystem bound-aries. Second, the number of partitions detected is also takeninto consideration where the greatest number of partitions issought because it provides the greatest granularity withinthe matrix. In terms of the CAD files, the greatest granularitywould provide the lowest level of subsystem within theproduct architecture. Third, the effect of pruning is takeninto account because increasing the pruning often leads to agreater loss of models due to disconnection from the rest ofthe matrix. Hence, it is desirable to ensure that a large propor-tion of the original matrix remains after pruning; thus, theDSM is as representative of the entire product architectureas possible. Fourth, the number of matrix components pro-duced by the pruning is considered. Here, fewer components

is preferable because components are disconnected from oneanother, and thus, there is a potential loss in understanding ofthe dependency between these components, yet it is also anindicator for the number of separate modules within a prod-uct’s design.

To achieve this, a range of time periods and levels of prun-ing are considered exhaustively, and the modularity, numberof communities, remaining network size, and number ofcomponents are determined. Figure 8 shows the results foreach of these parameters where the dashed rectangles high-light the desired areas for each of the parameters. The effectof sensitivity with respect to the four matrix metrics are thefollowing:

1. The partition modularity (Fig. 8a) has a positive corre-lation between the time period and level of pruning inorder to sustain a high modularity score.

Fig. 8. Effect of time period and pruning on the community partition analysis.




https://doi.org/10.1017/S0890060416000391


2. The optimum number of communities is more sensitiveto both the level of pruning and time period with a max-imum attained at 0.5 h and level of pruning of 0.3(Fig. 8b).

3. Beyond a pruning level of 0.35, the remaining numberof models within the DSM decreases (Fig. 8c).

4. The number of components (Fig. 8d) follows a similarpattern to the number of communities whereby it is bothhighly susceptible to changes in the time period andlevel of pruning.

Because the ideal time period and level of pruning differsfor each of the parameters, it can be seen that a compromisemust be made. In addition, black areas are a result of the entireremoval of the matrix by pruning.

To achieve this, the four metrics are combined as shown inEquation (2) to determine an overall measure O, where M isthe modularity of the matrix partition, Np is the number of ma-trix partitions, R is ratio of the network that remains afterpruning, and Nc is the number of matrix components. Theaim is to achieve the highest O, given a range of time periodsand pruning levels.

O ¼ MNpR

Nc: ð2Þ

Figure 9 shows the result of this search for an optimal set ofparameters and reveals a number of regions, such as 1 h and a0.25 level of pruning and 6.5 h and a 0.35 level of pruning. Inthis paper, it has been decided that 1 h and a pruning level of0.25 is the most suitable parameter set. For the purpose of this

study, the four matrix assessment metrics have not beenweighted, but it may be desirable to do so.

3.8. Summary

The previous sections have presented a detailed descriptionof the seven-stage process undertaken to generate the DSMsalongside an example of generating a DSM for the CADmodels from the Formula Student data set. Althoughmuch of the process can be automated, a number of deci-sions have to be made given the context of the data setthat is to be analyzed. First, the initial model filtering wasbased upon activity. In this paper, this was specified as aminimum of four updates. Second, it is also necessary to de-cide whether to continue with a directed or undirected co-occurrence matrix. Through investigation of the one- andtwo-way co-occurrence values, an undirected matrix waschosen. Third, this decision concerns the time period andlevel of pruning of the matrix. In order to make a reasoneddecision on these values, the process introduces a searchstage where these values can be chosen based on four mea-sures: the modularity, the number of partitions, the number ofcomponents, and the remaining matrix size. The sensitivityof these parameters has been discussed based on equalweightings of the four measures. This revealed a suitabletime of 1 h and pruning level of 0.25.

4. RESULTS AND DISCUSSION

As previously discussed in Section 2, the data set consideredin this study involves the design of a Formula Student car and

Fig. 9. Determining the appropriate time period and pruning values.




https://doi.org/10.1017/S0890060416000391


the evolution of the associated CAD, CFD, FEA, and WAVEmodels. Correspondingly, three DSMs are constructed usingthe process described in Section 3. The first DSM is a CADmodel only DSM. The second comprises CAD, CFD, FEA,and WAVE product models, with the OptimumK modelsnot being considered due to a lack of file activity being re-corded. The third DSM is a dynamic product model DSMthat captures the evolution of dependencies across all productmodels in real time as the project progresses from week toweek. The following discussion centers on the validity ofthe DSMs through comparison with the existing product ar-chitecture as defined by the team and results from previousDSM studies, and the additional insights that could bebrought by an automatic DSM process.

4.1. CAD DSM

The properties of the CAD DSM are summarized in Table 3.The table shows that a high proportion of the models remainafter the pruning stage (85.5%) with a core number of co-occurrences remaining (2.83%). This highlights that a signif-

icant number of candidate dependencies are likely to havebeen false positives. The modularity score of 0.85 highlightsthat a high level of structure exists within the matrix and sug-gests that the product architecture can be clearly separatedinto a number of subsystems. In addition, the high numberof components (13) demonstrates that the products design isalso highly modular in parts, an insight that would be difficultto ascertain through the existing hierarchical structure of theCAD models as the dependencies could span system and sub-system boundaries. However, a total of 12 additional parti-tions were generated through Louvain partitioning, whichhighlights that there are a number of highly integrative sub-systems within the product architecture.

Figure 10 shows the impact of the partitioning on the DSMstructure. Initially, it appears that there is no structure to theDSM (Fig. 10a), but postpartitioning, a high-level of structureis shown to exist (cf. Fig. 10b). This high level of structurewithin the generated DSM is comparable to the DSMs gener-ated by Sosa et al. (2003) and Gorbea et al. (2008) for a jetengine and hybrid vehicle, respectively. This result is alsoin line with manually curated DSMs that have been producedby Van Beek et al. (2010) and Maurer (2007) within the For-mula Student context.

It has been noted by Eppinger and Browning (2012) thatone of the key benefits of DSM is the ability to visualizseand interrogate subsystem dependencies through figuressuch as Figure 10b. However, it is also noted that this DSMcan be considered very large in terms of the number of ele-ments being visualized when compared to more traditionalDSMs, which commonly feature between 10 and 100 matrixelements (Eppinger & Browning, 2012) Consequentially,such a matrix might require additional support in terms of na-vigation in order for an engineer to comprehend all the depen-dencies but also demonstrates the greater level of granularityafforded by the automatic process.

Table 3. Summary of CADanalysis

Description Value

Undirected matrixModels 293Co-occurrences 21,679

Pruned matrixModels 253 (85.5%)Co-occurrences 613 (2.83%)Components 13Partitions 25Modularity 0.850

Fig. 10. Community partitioning on the co-occurrence of computer-aided design models.




https://doi.org/10.1017/S0890060416000391


It is immediately apparent that the partitions generatedfrom the co-occurrence of file activity have identified manydependencies that are cross system and subsystem. Therefore,it is argued that the analysis provides further insights into thedesign of the Formula Student car that the current CAD man-agement system does not provide. While BAPs manage andmonitor the product architecture, the automatic DSM processoffers potential for identifying dependencies beyond purelyarchitectural (Figure 11). These partitions could relate tofunctional, structural, and/or process dependencies.

Focusing on the partitions generated by the automated DSMprocess, four key insights can be drawn. A more complete over-view of the partition breakdown is presented in Table 4.

1. Areas of the product exist that are highly interdependentacross all the team-defined subsystems. Hence, achange in one of these models could require significantrework across many subsystems and might be an areathat requires careful monitoring, particularly in laterstage design.

2. There are a number of smaller partitions that consist ofonly one team-defined subsystem. Thus, a change inone of these partitions may only require changes to asmall subset of parts in the product and potentially anengineer from a single discipline.

3. There are a number of partitions that arise from one sub-system, and this may indicate that the subsystem can be

Table 4. Summary of CAD analysis

Partition Dependency Interpretation

0 Functional dependency between the steering system and suspension system, CAD models may influence thesteering and dynamic performance of the car

1, 7, 12, 17, & 25 Interdependent frame and body system, a potential single module of product2, 4, & 5 Dependency between frame and body, potential structural interfaces between systems3 Dependency between engine and drivetrain and frame and body components, potential engine fixtures

locations with respect to the frame6 Engine and drivetrain module with interfaces with some standard components8 & 22 Interdependent steering system partition, potential single module of the product9 Engine and drivetrain, suspension system, and wheels dependency, potential functional dependency

affecting dynamic vehicle performance (e.g., body role)10, 14, 16, 18, 21, & 23 Interdependent engine and drivetrain or wheel system, potential single module of product13, 20 Interdependent standard parts partition, potential identification of standard parts from nonclassified parts15 Dependency between suspension and wheel systems, potential structural and functional dependency19 Interdependent braking system partition, potential single module of the product25 Dependency between frame and body systems, potential identification of standard parts from nonclassified

parts

Fig. 11. Comparison of bath automated parts system and design structure matrices structures.




https://doi.org/10.1017/S0890060416000391


further decomposed. For example, further decompositionmight be useful for the purpose of resource allocation.

4. The partitioning has identified distinct subsystem de-pendencies. Partition 3 is one such example, where spe-cific models of the Engine & Drivetrain are related tomodels of the Frame & Body. This is a logical outcomebecause the engine fixtures are located to specific areasof the frame.

From this evaluation, it is argued that the DSM provides in-sights that are consistent with the known structure of the For-mula Student vehicle, and the results are consistent withDSMs generated from more traditional approaches. In addi-tion, a number of additional benefits have been revealed byusing an automatic DSM generation process. Therefore, thisprovides confidence in the proposed approach to produceDSMs.

The partitioning of the matrix also enables the models to bemeaningfully grouped by the intensity of the dependencies.This is in effect moving up a level in the hierarchy of theproduct architecture, and the results are shown in Figure 12.The rows and columns represent the partitions from the pre-vious matrix (Fig. 10), and the color provides a visual indica-tion of the level of dependency between the partitions. Thematrix is symmetric as it remains undirected.

It can be seen in Figure 12 that partitions beyond 15 appearto not have dependencies with the other partitions, and this isdue to the fact that these are small components that were gen-erated by the pruning of the matrix, and the potential depen-dencies were removed as false positives. Partitions 1 to 9 doshow dependencies between one another, and in particular,there appears to be a high dependency between Partitions 3

and 5, and 5 and 9. This suggests, that Partition 5 is a funda-mental integrating partition of CAD models and, hence, likelyto be a core subsystem of the product, which in this case is theFrame & Body. A change in this partition is likely to have asubstantial impact across many other partitions within theDSM.

4.2. Product model DSM

Similarly to the CAD DSM, the properties of the productmodel DSM are given in Table 5. Consistent with the CADDSM, the pruning of the matrix achieved a greatly reducednumber of dependency candidates (80%) while maintaininga high proportion of the original matrix (75%). Again, themodularity value of 0.769 is comparable with the CADDSM with 30 components being generated and 45 partitions.

Figure 13 shows the results from the partitioning of the ma-trix using the DSM process. Similar to results shown in the

Fig. 12. Induced matrix from grouping the partitioned computer-aided design models.

Table 5. Summary of CAD, CFD,FEA, and WAVE DSM

Description Value

Undirected matrixModels 488Co-occurrences 51,245

Pruned matrixModels 367 (75.3%)Co-occurrences 1,020 (1.99%)Components 30Partitions 45Modularity 0.769




https://doi.org/10.1017/S0890060416000391


CAD DSM (Fig. 10), Figure 13a shows a seemingly randommatrix while the partitioning result in Figure 13b reveals thatthere is a significant level of structure within the productmodel DSM. However, in contrast to the CAD DSM, thereappears to be greater interdependency between the partitions,which is shown by an increase in co-occurrences that existoutside the partitions denoted by the blue squares.

To investigate this result further, the model types have beenmapped to the partitions generated from the DSM. Figure 11shows the results of this mapping with four interesting fea-tures being exposed.

1. Each model type has a partition that consists solely of itsrespective model. This is a logical output because manyof the components within the product may not requirean in-depth analysis in relation to CFD, FEA, orWAVE modeling.

2. The existence of a single WAVE model partition is un-related to the rest of the other modeling types. BecauseWAVE models involve the analysis of the internal com-bustion parameters of the internal combustion engine,the modeling looks to optimize the engine mapping ra-ther then look to change the design of the engine itself.Therefore, this is confirmation that the DSM has theability to detect this as a separate partition and furtherprovides evidence of the technique’s validity. This ismay also be the case for some of the CFD models wherethey may be testing new components that did not be-come part of the final Formula Student car.

3. Identification of dependencies between areas of theCAD model and the FEA and CFD models in thisanalysis highlights that a change in any of these modelsmay require further FEA and CFD analysis to be per-formed.

4. It is argued that a single partition consisting of CAD,CFD, and FEA models may be an area that will require

monitoring by project management because any changewith these models may lead to considerable reworkacross multiple engineering disciplines

As with the CAD DSM, the product model DSM can use thepartitions to group the models and form a secondary matrix ofthese partitions. Figure 14 shows the results of this inducedproduct model matrix, and the result is similar to that of theCAD DSM induced matrix. Above Partition 16, there areno dependencies observed between the partitions becausethese are the components that have been generated from Prun-ing 16, and below Partition 16 show the inner partitions of thelarger components within the network, and it can be seen thatthere exists potential dependencies between these partitions.In particular, Partition 16 shows the greatest number of de-pendencies between the other partitions, and looking backto Figure 15, it can be seen that this partition is the largestin terms of number of models, and this may highlight thekey interfacing components within the product.

4.3. Dynamic network analysis of models

The final aspect that has been investigated is the potential ofautomatically generating DSMs during the engineering proj-ect. This has been achieved through the generation of productmodel DSMs based on the weekly model activity during theFormula Student project.

Figure 16 provides a summary of the matrix statistics andhow they evolve throughout the weeks of the project. It is im-mediately apparent that a gap exists within the updates of theproduct models between Weeks 7 and 8, and this is due tothe students going for Easter vacation and work ceasing onthe project. Figure 16a shows that there is a consistency inthe level of work being performed on the CAD, CFD, andFEA models, while the WAVE model activity only appearsbetween Weeks 4 and 6. Figure 16b shows the statistics for

Fig. 13. Community partitioning on the co-occurrence of product models.




https://doi.org/10.1017/S0890060416000391


the DSMs created for each week. An interesting feature of theDSMs is the variability in the size of the matrix between theweeks and modularity score for the partitioning. It appearsthat there is a limit in generating “good” partitions with lowmodel activity. This is a logical result, and the lack of modelco-occurrence activity makes the identification and distinc-tion between positive and false positive dependencies morechallenging. It could also be more difficult to detect subsys-tem structures on a weekly basis because the team are focus-ing on different elements of the product architecture. Thus,the partitions may be more reflective of the areas of work ra-

ther than identifying a particular product subsystem as withthe previous CAD and product model DSMs.

Figure 17 shows the DSMs for Weeks 2 to 7 of the FormulaStudent project. The challenges of low model activity and theability to generate a DSM is more apparent when one com-pares Figures 17c and 17d. Figure 17c clearly has fewer activemodels, and if it were to be compared to the entire DSM fromSection 4.2, this may show the mode of working of the teamas well as the specific area of work. In addition, comparingthese DSMs to the previous CAD and product modelDSMs, it can be seen that there appears to be a great deal

Fig. 15. Composition of product model design structure matrices partitions based on model type.

Fig. 14. Induced matrix from grouping the partitioned product model design structure matrices.




https://doi.org/10.1017/S0890060416000391


more interdependencies between the partitions that are gener-ated. This may be because the analysis is focusing upon aweek of activity where it is suggested that the team are focus-ing on integrating models within a few areas of the product ata time. Because the analysis has used the same time periodand pruning for the consistency across the analysis, it maybe the case that these values may not be suitable for the anal-ysis of individual weeks, and hence, there may be more falsepositive within these DSMs.

Figure 18 provides further details on the composition of thepartitions from Weeks 2 to 7 of the Formula Student project.From comparison of these weekly compositions, four inter-esting features emerge.

1. Week 2 (Fig. 18a) reveals partitions consisting of modelsfrom all the subsystems in the product. In particular, Par-tition 1 focuses on the Suspension, Frame & Body andSteering System, so it is suggested that these componentsare forming the initial chassis for the car. Moving to Par-tition 2, this solely consists of standard parts and parts thatcould not be classified through their naming convention.This provides evidence to suggest that the nonclassifiedparts are part of the standard library of parts of the car.Partition 2 consists of parts from the rest of the subsys-tems for the car, and this could be the initial placeholdersfor the arrangement of the subsystems within the car.

2. Concentrated working on particular subsystems hasbeen identified in Week 3 (Fig. 18b) and Week 4(Fig. 18c) though the majority of the partitions consistof one model type. This may also indicate that thesesubsystems require a certain level of maturity beforebeing introduced to the rest of the product architecture.

3. Partition 1 in Week 4 (Fig. 18c) highlights a particulararea of multidisciplinary collaboration between theFrame & Body development of the CAD alongsidethe associated CFD and CFD and FEA analysis.

4. A change in the mode of working within the teamemerges after Week 4. Weeks 3 and 4 (Fig. 18b,c) focusedon specific areas alongside CFD and FEA analysis;Weeks 4–7 (Fig. 18d–f) show a high-level of integrativeworking between the various subsystems of the car. Ini-tially, the majority of partitions within Weeks 4 and 5(Fig. 18c,d) consist of either two or three different subsys-tems, which could be used to identify specific dependen-cies between subsystems. The level of integrative workingcontinues to grow in Week 7 where both the number ofpartitions and number of subsystems within each partitionincrease. This shows that as the project and product de-velop, the more integrative the work becomes.

These results highlight the potential of automatically generat-ing DSMs as engineering projects progress and the potentialinsights that they could provide to support engineering proj-ect management and the design process. In addition, a betterunderstanding of how the multitude of dependencies evolveand develop within a product could also be produced, andhaving a library of DSMs from past product developmentscould be used to support and monitor the development andprogression of new products.

4.3. Summary

The results from DSMs generated from the Formula Studentproject has provided considerable evidence to suggest thatthe process of generating DSMs through the co-occurrenceof product model updates provides valid and useful informa-tion for engineering project management. In addition, the re-sults show the potential for this process to produce dynamicproduct model DSMs, which would have previously been atime consuming and potentially costly endeavor. Table 6 pro-vides a summary of the key features that have been identifiedthrough the generation of the CAD, product model, and dy-namic product model DSMs.

Fig. 16. Model activity and design structure matrices statistics for the dynamic design structure matrices.




https://doi.org/10.1017/S0890060416000391


Fig. 17. Partitioning results for Weeks 2 and 7 of the formula student project.




https://doi.org/10.1017/S0890060416000391


5. FUTURE WORK

In generating the process and interpreting the results, a numberof avenues have been revealed for potential future work. The

first avenue lies in applying this technique within an industrialcontext as a current limitation of this study is that the data set isone of a university-based project. Two challenges are foreseen.The first is that for the study reported in this paper, the team

Fig. 18. Partitioning composition for Weeks 2 and 7 of the formula student project.




https://doi.org/10.1017/S0890060416000391


was colocated, and thus, the co-occurrence of work was per-formed within the same time zone. In contrast, large engineer-ing projects may potentially have many concurrent work pack-ages that are being worked upon in a distributed manner acrossmultiple time zones. The second concerns the time it takes forchanges to occur within various product models. For example,small CAD changes can take within the region of 2 min whileCFD analyses could take days to run and complete. Thus, CADmodels are open to more co-occurrence events than CFD,which the analysis has yet to adjust for.

Although there remain a few limitations, it is also arguedthat the analysis in its current form could support small to me-dium enterprises where these challenges may not be manifestand the cost of implementing PLM systems is too great. In ad-dition, it would be interesting to compare this process along-side existing DSMs practices within engineering companiesas well as their change management processes.

The second avenue of future research lies in the delivery ofthe results. As discussed previously in Section 4.1, the DSMsgenerated by this technique are much larger than more tradi-tional DSMs, where the number of elements are typically inthe hundreds, while this technique could produce DSMswith thousands of elements. Therefore, novel methods in in-teracting and navigating these DSMs to support engineers intheir understanding of the multitude of dependencies thathave been identified would be interesting to explore further.

The third is in the assessment of the impact of providing thisinformation into Formula Student projects in future years. Aninteresting research question is to understand how the resultsof this analysis could be used to inform and support future en-gineering projects. In addition, expert evaluation of the DSMsgenerated through this process could be used to ascertainwhether the dependency is functional, structural, or procedural.With this additional information, future work could seek to au-tomatically determine and further verify the type of dependencyalongside the existence of a dependency as the engineering proj-ect evolves. The Formula Student projects also provides a con-sistent project and engineering process where DSMs can begenerated year on year, and it would be interesting to investigatethe consistency of the DSM from team to team to explore howteams performing the same task impact the DSM.

The fourth and final avenue concerns whether the type andlocality of the dependency can be uncovered through the anal-ysis of the changes in content between product models. In ad-dition, the dynamic DSM analysis has focused on the timeslices of engineering activity, and analysis of the cumulativeactivity may provide additional insights in terms of the evolu-tion of the dependency structure as well as the engineeringprocess that has been followed.

6. CONCLUSION

Managing component interactions and dependencies remainsa core and fundamental element of engineering and is facedby engineers on an almost daily basis. Being able to detectand monitor these interactions and dependencies continuesto be a challenging area for engineering project management,and one that is set to become even more challenging as thenumber of components, component interactions, and compo-nent dependencies of products is continually increasing.

This paper has sought to complement existing techniques byextending the well-established DSM method through the crea-tion of a process that automatically generates DSMs from theevolution of product models. This process has been appliedto the evolution of the product models associated with a For-mula Student project in order to assess the validity and additioninsights that could be brought by an automatic DSM process.

Three DSMs were generated from the Formula Studentdata set, a CAD DSM, a product model DSM, and a dynamicproduct model DSM. The results reveal that the process isable to produce a DSM that is representative of the vehiclewhen compared to the existing subsystem definition and iscomparable to DSMs generated by traditional methods. In ad-dition, areas of high dependency and isolated subsystems ofthe vehicle have been identified, and this information couldbe used to support resource allocation and change manage-ment within the project. Finally, the key affordance offeredby the process is the ability to provide dynamic DSMs inreal time to an engineering project. This would be previouslyimpractical using traditional DSM generation practices, andthe results have shown that they can indicate changes in the

Table 6. Summary of CAD, CFD, FEA, and WAVE DSM

CAD DSM

1. Areas of the product exist that are highly interdependent across all thesubsystems that have been defined.

2. There a number of smaller partitions that consist of only one subsystem.3. There are a number of partitions that arise from one subsystem and this

may indicate that the subsystem can be usefully further decomposed.4. The partitioning has identified distinct subsystem dependencies.

Product Model DSM

1. Each model type has partitions that solely consist of their respectivemodels.

2. The existence of a single WAVE model partition, which is unrelated to therest of the other modeling types

3. Identification of dependencies between areas of the CAD model and theFEA and CFD models where this analysis highlights that a change in anyof these models may require further FEA and CFD analysis to beperformed

4. A single partition consisting of CAD, CFD, and FEA models

Dynamic Product Model DSM

1. Partitions in Week 2 consist of models from all subsystems in the productthat may indicate the generation of the initial layout of the vehicle

2. A change in the mode of working in Weeks 3 and 4 indicatingconcentrated work on areas of the vehicle

3. Identification of areas in the product architecture that requiremultidisciplinary collaborative work

4. A change in the mode of working that indicates a high-level of integrativeworking between the various subsystems of the vehicle




https://doi.org/10.1017/S0890060416000391


mode of working within the project and provide insights intohow dependencies between subsystems have emerged.

ACKNOWLEDGMENTS

Thework reported in this paper was undertaken as part of the Languageof Collaborative Manufacturing Project at the University of Bath andthe University of Bristol, which is funded by Engineering and PhysicalSciences Research Council (EPSRC) Grant EP/K014196/2.

REFERENCES

ANSYS. (2015). Ansys cfx. Accessed at http://www.ansys.com/Products/Simulation+Technology/Fluid+Dynamics/Fluid+Dynamics+Products/ANSYS+CFX

Blondel, V.D., Guillaume, J.-L., Lambiotte, R., & Lefebvre, E. (2008). Fastunfolding of communities in large networks. Journal of Statistical Me-chanics: Theory and Experiment 2008(10), P10008.

Briggs, D. (2012). Establish digital product development (dpd) low endviewer (lev) and archival standard for 787 project. Proc. Collaboration& Interoperability Congr. (CIC), Denver, CO, May 21–23.

Browning, T.R. (2009). Using the design structure matrix to design program or-ganizations. Handbook of Systems Engineering and Management (Sage,A., & Rouse, W., Eds.), 2nd ed., pp. 1401–1424. Hoboken, NJ: Wiley.

Bruce, C. (2015). Toyota recalls 625k Prius models for faulty hybrid soft-ware. Accessed at http://www.autoblog.com/2015/07/15/toyota-recalls-625k-prius-faulty-hybrid-software/

Calleam. (2011). Airbus—a380. Accessed at http://calleam.com/WTPF/?p=4700Chen, S.-J.G., & Huang, E. (2007). A systematic approach for supply chain

improvement using design structure matrix. Journal of Intelligent Manu-facturing 18(2), 285–299.

Eppinger, S.D. (1997). A planning method for integration of large-scale en-gineering systems. Proc. Int. Conf. Engineering Design, pp. 199–204,Tampere, Finland.

Eppinger, S.D., & Browning, T.R. (2012). Design Structure Matrix Methodsand Applications. Cambridge, MA: MIT Press.

Fixson, S.K. (2005). Product architecture assessment: a tool to link product,process, and supply chain design decisions. Journal of Operations Man-agement 23(3), 345–369.

Gorbea, C., Spielmannleitner, T., Lindemann, U., & Fricke, E. (2008). Anal-ysis of hybrid vehicle architectures using multiple domain matrices. DSM2008: Proc. 10th Int. DSM Conf., pp. 375–387, Stockholm, November11–12.

Hagberg, A.A., Schult, D.A., & Swart, P.J. (2008). Exploring network struc-ture, dynamics, and function using NetworkX. Proc. 7th Python in Sci-ence Conf. (SciPy2008), pp. 11–15. Pasadena, CA, August 11–15.

Hamraz, B., & Clarkson, J. (2015). Industrial evaluation of fbs linkage amethod to support engineering change management. Journal of Engi-neering Design 26(1–3), 24–47.

Jarratt, T.A.W., Eckert, C.M., Caldwell, N.H.M., &, P.J. (2011). Engineeringchange: an overview and perspective on the literature. Research in Engi-neering Design 22(2), 103–124.

MacCormack, A., Rusnak, J., & Baldwin, C.Y. (2006). Exploring the struc-ture of complex software designs: an empirical study of open source andproprietary code. Management Science 52(7), 1015–1030.

Maurer, M.S. (2007). Structural awareness in complex product design. PhDThesis, University of Munich.

Newman, M.E.J. (2004). Analysis of weighted networks. Physical Review E70(5), 056131.

Newman, M.E.J. (2006). Modularity and community structure in networks.Proceedings of the National Academy of Sciences 103(23), 8577–8582.

Pimmler, T.U., & Eppinger, S.D. (1994). Integration Analysis of Product De-compositions. Cambridge, MA: MIT Press.

Pujol, J.M., Erramilli, V., & Rodriguez, P. (2009). Divide and conquer: par-titioning online social networks. Accessed at http://arxiv.org/abs/0905.4918

Raspberry Pi Foundation. (2015). Raspberry pi—home page. Accessed athttp://www.raspberrypi.org

Ricardo. (2015). Wave—1d engine gas dynamics, 2015. Accessed at http://www.ricardo.com/en-GB/What-we-do/Software/Products/WAVE/

Senescu, R.R., Head, A.W., Steinert, M., & Fischer, M.A. (2012). Generatinga network of information dependencies automatically. Proc.14th Int. De-pendency and Structure Modelling Conf., DSM12, pp. 139–152, Stan-ford, CA, September 12–13.

Siemens. (2015). Nx: Siemens plm software. Accessed at http://www.plm.automation.siemens.com/en\_us/products/nx/

Sosa, M.E., Eppinger, S.E., & Rowles, C.M. (2003). Identifying modular andintegrative systems and their impact on design team interactions. Journalof Mechanical Design 125(2), 240–252.

Sosa, M.E., Eppinger, S.D., & Rowles, C.M. (2007). A network approach todefine modularity of components in complex products. Journal of Me-chanical Design 129(11), 1118–1129.

Steward, D.V. (1981). The design structure system: a method for managingthe design of complex systems. IEEE Transactions on Engineering Man-agement 28(3), 71–74.

Van Beek, T.J., Erden, M.S., & Tomiyama, T. (2010). Modular design ofmechatronic systems with function modelling. Mechatronics 20(8),850–863.

James Gopsill is a Senior Research and Teaching Associatein the Department of Mechanical Engineering at the Univer-sity of Bristol. As a member of the Design & ManufacturingFutures Lab, his research interests cover the application of ar-tificial intelligence, machine learning, and big data analyticsto support engineering design and knowledge management.In addition, he has an interest in rapid prototyping technolo-gies and the development of tools to support their applicationwithin the engineering design process. He has publishedmore than 25 journal and conference papers and has receivedtwo outstanding contribution awards.

Chris Snider is a Senior Research and Teaching Associate inthe Department of Mechanical Engineering at the Universityof Bristol. Working within the engineering design and man-ufacture group and extensively with industry, his research in-terests cover development of engineering design processes,design support tools and methods, design thinking and be-havior, engineering informatics, virtual and physical proto-typing, and engineering management. In particular, his re-search focuses on monitoring and analysis of engineeringprocesses throughout their life cycle, as well as effective de-sign and development within constrained design situations.He has published more than 25 journal and conference papersand has won three international best paper awards.

Chris McMahon is a Professor of engineering design in the De-partment of Mechanical Engineering at the Universityof Bristol, apost he has held since September 2012. He previously worked atthe University from 1984 to 2002. From 2002 to 2012, he workedat the University of Bath as Reader and then Professor and Direc-tor of its Innovative Design and Manufacturing Research Centre.Prior to 1984 he was a production and design engineer in the rail-way and automotive industries. His research interests are in engi-neering design, especially concerning the application of compu-ters to the management of information and uncertainty in design,design automation, product life cycle management, design educa-tion, and design for sustainability, areas in which he has publishedover 250 refereed papers, a textbook, and a number of editedbooks. Professor McMahon is a Chartered Engineer, Fellow ofthe Institution of Mechanical Engineers (UK), and a founder



http://www.ansys.com/Products/Simulation+Technology/Fluid+Dynamics/Fluid+Dynamics+Products/ANSYS+CFX




http://www.autoblog.com/2015/07/15/toyota-recalls-625k-prius-faulty-hybrid-software/



http://calleam.com/WTPF/?p=4700

http://calleam.com/WTPF/?p=4700

http://arxiv.org/abs/0905.4918



http://www.raspberrypi.org

http://www.raspberrypi.org

http://www.ricardo.com/en-GB/What-we-do/Software/Products/WAVE/



http://www.plm.automation.siemens.com/en\_us/products/nx/




https://doi.org/10.1017/S0890060416000391


member of the Design Society, for which he was President from2010 to 2013. He is an active memberof the scientific committeesof various international journals and conferences.

Ben Hicks is a Chartered Engineering and Professor of me-chanical engineering at the University of Bristol. He isHead of Engineering Systems and Design and leads the De-sign and Manufacturing Futures Lab. His expertise spans ma-

chine and manufacturing systems design, including self-repli-cating machines, modeling machine–material interaction,virtual prototyping, and engineering informatics. Dr. Hickshas published over 190 journal and conference articles, is apast member of EPSRC’s Strategic Advisory Team, and recip-ient of the IMechE Water Arbitration Prize for best originalpaper.




https://doi.org/10.1017/S0890060416000391


automatic generation of design structure matrices through the … · 2017-11-28 · faculty of...

Documents